What about these typical songs? A closer look at some track-level features

---

Here we see the typical songs for each artist as mentioned earlier depicted by four track-level features. Some of the things I noticed myself are also representedin the graph.

The Fearless Flyers have the highest energy by far, while the love songs of Theo Katzman and the minimalist funk of Vulfpeck get low scores.

For Theo Katzmans song the valence is also the lowest, which often corresponds to sad songs, much like the melancholy often felt when listening to love songs.

The Dancability feature then seems odd in combination with the others. I would have expected Cory Wong to be the highest by far in this category.

Cory Wong scores high in Valence however, perhaps it was my personal correlation between dancing and happiness, which is often said to be represented by a combination of valence and energy in music.

Finally, Nate Smith scores lowest in Liveness. The track used is a solo drum track, which might explain the low score due to a lack of noise in the recording.

We are certainly gaining some insight into the differences between the artists, but these are just some typical tracks. Lets look at the bigger picture!

Lets see what emotions best describes each artist

---

Here we see the entire dataset plotted based on valence on the x-axis and energy on the y-axis.

We can already begin to see some larger patterns. Most of the music is happy, and you can see that Fearless Flyers music actually never leaves the happy quadrant. Cory Wongs music also rarely does, but sometimes has a lower valence and ends up in theh angry quadrant.

You can clearly see Theo Katzman, Vulfpeck and Woody Goss are on the lower side of the graph. Overall they have less energy. For Theo Katzmann this is because of the style of music he usually plays, which is half-time smooth love songs. Woody Goss’ and Vulfpeck’s reason is most likely the minimalist production they use. Short notes and interesting rests in the music are something that makes their music uniqie.

Nate Smith is the only one who shows a clear pattern of low valence and high energy. This somewhat understandable, a lot of his music features heavy drum parts which in audio analysis often leads to high and intense peaks. It is difficult to descibe how one would express valence in terms of drums.

Major and Minor modes are often associated with happy and sad, perhaps that is the explanation?

---

Tempo and Loundess seem to have little correlation. We can however see that for most artists songs that are high in both tempo and loudness have a higher chance of being in a major key.

Combining Mode with Key

---

Although the distribution is clearly not random, the modes used don’t seem to correspond with the, perhaps too simple, assumption that we can explain the emotion of the music using just the mode.

We can clearly see Theo Katzman uses almost exclusively Major keys, but still his songs were relatively low energy and valence compared to the other groups, which feature a lot more Minor keys.

Cory Wong, whose music I find the most uplifting of all, has relatively many minor keys.

One thing that is striking is the identical gap in keys between Vulfpeck and the Fearless Flyers for the keys of D# and E. Both these bands feature Joe Dart, perhaps he doesn’t like those keys very much?

Funnily enough, in his solo repetoire Cory Wong also doesn’t prefer d#, but does have a few songs in E.

Different genres have different tempi, what’s our dataset like?

---

We see the mode tempo for this dataset is 106 BPM. Compared to popular music, whose average tempo has risen up to 120 BPM in recent years, this is on the slower side.

There are also smaller peaks around 70, 90, 130, 155 and 170 BPM.

We can see that the graph is slightly left skewed which indicates that lower tempos are more constant, while higher tempos are more variable. .

.

.

.

.

.

And how do the categories differ?

---

We can start to see Tempo might be a pretty good predictor in certain cases. Nate Smiths Music has a clear peak around 90 BPM. The Fearless Flyers also havea high tempo as their mode, around 130 BPM, which corresponds to my expectation mentioned earlier. This also explains the peak in the earlier graph.

Although there seem to be clear areasfor each plot, we can see the means are very close together. This is likely caused due to the outliers you can see in almost every group except for the Fearless Flyers. .

.

.

.

.

Vulfpeck - 1612

---

0-20 seconds

The song opens with a bass melody that continous throughout the song.

100-110 seconds

A vocal solo

140-140 seconds

Vocal fills

In the timbre-based self similarity matrix on the left we see a bright glow indicating that, overall, the timbre of the song is not very constant This is quite a surprise, the song features a bass and keyboard that can be heard throughout the song. Perhaps the combination of instruments and vocals create sounds that the analysis simply couldn’t keep apart

The pitch-based self similarity matrix tells the opposite story. It is dark throughout, indicating a steady pitch, except for two key moments, which are caused by the vocal breaks. The vocals on this track are highly improvised which explains the clear difference with the repetitive melodies of the other instruments.

It’s interesting to see that the vocals have such a high influence on the overall structure of the song

.

.

.

.

.

.

.

.

.

.

.

.

.

.

The Fearless Flyers - Introducing The Fearless Flyers

---

0-40 seconds Main theme

Drum fill

50-70 seconds Main theme again

Drum fill

80-110 seconds Extended build up

120-140 seconds Drum solo

The song here is clearly divided in 4 parts. The first two parts are very similar, which is also what you hear. The playing style is the same and this is reflected in the dark squares. Each part is divided by a drum fill which differs from the rest of the song. It features three string instruments at the same time. The final part of the song is a drum solo, it has little similarity with the rest of the track.

As for the melody, the same pattern is visible, but in a much more exagarated manner. The largest bright lines occur when there string section goes from fast strumming to a melodic build up with longer notes. The drums seem to have a high influence in the pitch similarity still as even the final part has a high similarity with the rest, even though it features little strings. .

.

.

.

.

.

.

.

.

.

.

.

.

.

Nate Smith- Big/Little Five

---

0-120 seconds Hi Hat grooves

120-150 seconds Ride Grooves

150-190 seconds Back to Hi Hat

Here we can really see the power of timbre analysis. Drums are a limited instrument in melodic terms, but the difference between different drums is something everyone can hear and in this case an amazing pattern occurs. The small lines are caused by hits on the crash cymbal, which is often used on the first count of a 4 or 8 bar groove. The clear dark squares are caused by the different cymbals used for the eigth notes played in the groove. The distinction between the smaller HiHat and the larger Ride creates a difference in pitch that is clearly picked up by the chromatic analysis.

As you might expect, the pitches don’t really say that much. He simply plays the drums for a solid three and a half minutes, so the only variation is between the different drums, but no clear pattern occurs there. Again the biggest brightness is due to the lack of cymbals and can be seen in the opening 10 seconds.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Cory Wong - Golden

---

0-60 seconds Verse and Chorus

60-120 seconds Verse and Chorus

120-180 seconds Verse and Chorus

180-240 seconds Verse and Chorus

The timbre analysis does not show a very clear structure. The third repetition features a more pronounced brass section, while the guitar takes a little break.

The most striking pattern is in the pitch features. The Bright lines are caused by breaks at the end of the chorus, going back into the verse. Here you can see this pattern of verse and chorus is repeated 4 times with similar sounds.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Theo Katzman - Love Is A Beautiful Thing

---

0-200 seconds Soothing and rythmic guitar progression with a repeating vocal theme

The timbre feature seems to show very little similarity throughout the song, however, this is likely caused by the fact that the song is actually extremely similar.

The pitch analysis confirms the repetition I claimed above. The guitar progrression is a 4 bar loop, and the vocal melody on top of it starts each repetition with the same theme: “Love is a beautiful thing”. This exact repetition causes the diagonal lines in the graph

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Woody Goss - Be There

---

0-70 verse-prechorus-chorus

70-140 verse-prechorus-chorus

140-250 seconds A sing-along type vocal melody that is repeated

The timbre features again give little insight. The final part features a trumpet, which doesn’t occur throughout the rest of the song.

The strange checkerboard pattern that can be seen twice in is caused by the way the melody is built. It is an interplay between a bass on the down-beat and a high guitar chord on the up-beat. The final part is a repetitive vocal melody made to sing along.

Lets put this to the test, we certainly found some differences, but are they distinctive enough?

           Truth
Prediction  Cory Wong Fearless Katzman Nate Vulfpeck Woody
  Cory Wong        29        4       2    7       21     1
  Fearless          1       13       0    4        4     0
  Katzman           2        0      21    3        8     6
  Nate              1        0       0   12        4     0
  Vulfpeck         15        4       5    2       30     3
  Woody             0        1       3    1        7    14

---

When constructing the KNN network one thing that drastically improved performance is using only the top 10 most influential features. Still there are some interesting observations to be made. Although this confusion matrix gives some insight, it is skewed towards the bigger categories like Vulfpeck and Cory wong, and it only shows numbers, and not a clear measure of performance.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Detailed analysis of performance

# A tibble: 6 x 3
  class     precision recall
  <fct>         <dbl>  <dbl>
1 Cory Wong     0.615  0.667
2 Fearless      0.667  0.545
3 Katzman       0.586  0.548
4 Nate          0.778  0.483
5 Vulfpeck      0.548  0.689
6 Woody         0.444  0.333

---

As with the KNN network we see in this Random Forest implementation the Fearless Flyers playlist is the most distinctive group in the corpus. .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Importance of features

---

This shows the most important distinguishing features between the different playlists. Loudness seems to be the key distinguishing feature, followed by instrumentalness and energy. The rest of the top 10 is mainly timbre components, while the key of the song doesn’t seem to contribute all that much to the predictions. .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Clustering

[1] 228

  Vulfpeck Cory Theo Fearless Woody Nate
1       15    7   17        0     9    4
2       23   12    0       14     0    2
3        1    1    0        8     0   10
4       13   17    2        0     4    9
5        4    0    0        0     4    4
6       18   11   12        0     7    0

---

Here we try clustering in the 6 categories we defined at the beginning of this portfolio. As you can see the groups are certainly not pure. Nate smith seems to concentrate in the upper right, but is also in the lower clusters. I will analyse the exact classifications more precisely in the final portfolio. .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Principal Components

---

In the previous slide we were shown the 6 clusters our random forest came up with. This is represented in our graph along two dimensions. In machine learning Principal Component Analysis is used to determine a linear combination of the features in our dataset that offer the highest rate of variability. In this circle we see the same two dimensions, but now plotted on top is the contribution of each of our features, the ones we know how to interpret. We can see that acousticness and energy are the biggest contributours for dimension 1, while c01 plays the largest role in dimension two. .

.

.

.

.

.

.

.

.

.

.

.

.

.